YOU HAVE CROSSED THE VOID.

Riipen Level UP · Remote

• Engineered an end-to-end AI content pipeline processing 13 live news sources and 8 FRED economic data series daily, reducing database round trips by 99% by replacing per-row INSERTs with bulk execute_values operations across the full scraper layer.
• Designed and deployed a 3-agent CrewAI system using OpenAI that automatically compiles weekly economic newsletters, scanning, analyzing, and publishing bilingual drafts every Friday with zero manual intervention.
• Developed a repeatable performance measurement tool scoring article quality, translation glossary adherence, and workflow reliability from live production logs, providing a quantifiable baseline for ongoing optimization.

Python · Flask · Pandas · React

• Developed backend API using Flask to process and filter movie data from a dataset featuring 9,000+ movies, TV shows, and documentaries.
• Implemented filtering logic with Pandas for multi-dimensional queries across genre, language, rating, release date, and mature content.

Python · Scikit-learn · Pandas · NumPy

• Developed an NFL game outcome prediction model that achieved 76.9% winner prediction accuracy across 13 playoff games, by training on 46,000+ play-by-play records with 16 engineered features spanning offensive EPA, defensive efficiency, and turnover metrics.
• Built a score prediction system that estimated final game scores within 5 points 50% of the time and within 10 points 77% of the time, by implementing dual Random Forest Regressors trained on aggregated team-level statistics from 272 regular season games.
• Correctly predicted the Super Bowl LX winner and both Conference Championship outcomes (3/3), using season-long performance metrics to evaluate matchup strength.

Next.js · TypeScript · React · MongoDB · REST APIs · Tailwind CSS

• Designed an agentic AI system using Gemini 2.0 Flash with function-calling that autonomously orchestrates 5–12 targeted web searches per query, dynamically adapting its data collection strategy based on domain classification of the input.
• Optimized pipeline throughput by approximately 40% by mapping the dependency graph across 8–15 model calls and parallelizing independent inference stages while sequencing dependent ones.
• Built a multi-stage inference pipeline that processes unstructured web data into 130+ structured fields using JSON-mode prompting, with programmatic validation layers enforcing logical constraints.
• Engineered an autonomous research agent that adapts its search strategy per idea by leveraging Gemini's function-calling to orchestrate targeted API queries across competitors, market trends, regulations, and user sentiment.